Applied Thermal Engineering 67 (2014) 507e519

Contents lists available at ScienceDirect

Applied Thermal Engineering

journal homepage: www.elsevier.com/locate/apthermeng

Review of modeling methods for HVAC systems

Abdul Afram, Farrokh Janabi-Shari*
Department of Mechanical and Industrial Engineering, Ryerson University, 350 Victoria Street, Toronto, ON M5B 2K3, Canada

h i g h l i g h t s

Reviewed the major HVAC modeling techniques reported in the recent literature.

Physics based, data driven and grey box models are reviewed in detail.

Discussed the weaknesses and strengths of different modeling methods.

Comparison based on the robustness, ease of tuning and number of parameters etc.

a r t i c l e i n f o a b s t r a c t

Article history: This work presents the literature review of the methods used to model the heating, ventilation, and air
Received 9 August 2013 conditioning (HVAC) systems. The model development is necessary for the study of the energy con-
Accepted 23 March 2014 sumption of HVAC systems. Models are also required to simulate the different supervisory and local loop
Available online 2 April 2014
control strategies to improve the energy consumption efciency. HVAC systems have complex structures
consisting of heat and mass transfer equipment such as chiller, boiler, heating/cooling coils, and supply
Keywords:
air ducts. HVAC systems also consist of several sensors and controllers for regulating the controllable
HVAC models review
variables such as zone temperature, supply air temperature, supply air fan speed, duct static pressure,
Data driven models
Physics based models
and chilled water temperature at their set-points. To predict the energy consumption by the HVAC
Grey box models systems accurately, one needs to model the individual components either from the measured data or
Performance comparison metrics based on the knowledge of the underlying physical phenomenon. This results in three broad classes of
the models known as data driven, physics based, and grey box models. In this paper, major data driven,
physics based, and grey box modeling techniques reported in the recent literature are reviewed.

2014 Elsevier Ltd. All rights reserved.

1. Introduction
Modeling of heating, ventilation, and air conditioning (HVAC)
systems is necessary for studying and regulation of energy con-
sumption and quality of indoor environment. Generally three types
of modeling approaches are used for HVAC systems. In the rst one,
known as the data driven (black box or inverse) approach, the
system performance data is collected under normal use or under a
specic test and a relationship is found between the input and
output variables using the mathematical techniques [e.g., statistical
regression and articial neural network (ANN)] [1]. The second type
is known as physics based (white box or forward) approach, in
which the system models are derived using the governing laws of
physics and the detailed knowledge of the underlying process. In
* Corresponding author. Tel.: 1 416 979 5000x7097; fax: 1 416 979 5265.
E-mail addresses: abdul.afram@ryerson.ca (A. Afram), fshari@ryerson.ca
(F. Janabi-Shari).
the third type, known as the grey box approach, the basic structure
of the model is formed by using the physics based methods and the
model parameters are determined by using the parameter esti-
mation algorithms on the measured data of the system. In order to
build the physics based models and to determine their parameters,
the detailed knowledge of the system and its processes is necessary.
On the other hand, the data driven models and their parameters
can be obtained with little or no knowledge of the system and its
processes. The grey box models benet from the qualities of both
the physics based and data driven models as they use physics based
methods for building the model structure and use system perfor-
mance data to estimate the model parameters. Physics based
models have very good generalization capabilities but suffer from
poor accuracy; whereas, the data driven models have very high
accuracy on the training data but suffer from generalization beyond
the training domain. Grey box models benet from the advantages
of the other two types, providing good generalization capabilities as
compared to the data driven models and better accuracy as
compared to the physics based models.

http://dx.doi.org/10.1016/j.applthermaleng.2014.03.055
1359-4311/
2014 Elsevier Ltd. All rights reserved.
508 A. Afram, F. Janabi-Shari / Applied Thermal Engineering 67 (2014) 507e519
In this paper, a review of the data driven, physics based and grey
box models is provided. Comprehensive modeling resources are
available. In addition to the conference proceedings and journals,
American society of heating, refrigerating and air-conditioning
engineers (ASHRAE) has published several handbooks for HVAC
system fundamentals [1], equipment details [2], and its applica-
tions [3]. ASHRAE has also produced energy calculation toolkits for
the system primary components (e.g., boiler, heat pump, and
chiller) [4] and secondary ones [e.g., ducts, pipes, valves, dampers,
pumps, fans, and air handling unit or air handling unit (AHU)] [5]
which both include comprehensive dynamic, steady-state, and
quasi-steady state models. Also technical documentation of build-
ing energy simulation programs (i.e., TRNSYS [6], EnergyPlus [7],
and HVACSIM [8]) provide the models of common HVAC system
components. U.S. Department of Energy maintains a comprehen-
sive listing of building energy simulation software tools which in-
cludes popular tools like BLAST, BSim, ESP-r, and DOE-2. Several
toolboxes to simulate the building and HVAC systems have also
been developed for Matlab Simulink, e.g., international building
physics toolbox (IBPT) [9], SIMBAD toolbox [10], conventional and
renewable energy optimization toolbox (CARNOT) [11], HAM-tools
[12,13], and ASTECCA toolkit [14,15].
A few survey works focusing on some classes of modeling ap-
proaches have been reported in the literature. Examples include a
survey of fuzzy logic (FL) modeling for HVAC systems [16], software
used for the simulation of HVAC components [17,18], and Matlab
Simulink use in HVAC systems research [19]. While signicant
progress has been made in modeling HVAC systems, to the best of
the authors knowledge, no recent comprehensive survey of HVAC
systems models covering data-driven, physics based, and grey box
modeling approaches have been reported in the literature.
2. HVAC modeling techniques
HVAC modeling techniques are divided into data driven, physics
based, and grey box models. Models can also be classied as linear
or nonlinear, static or dynamic, explicit or implicit, discrete or
continuous, deterministic or probabilistic, and deductive, inductive
or oating models. According to this classication, most of the
physics based techniques fall under the deductive models whereas
data-driven techniques fall under the inductive models. Grey-box
models are hybrid models and can be categorized under both the
inductive and deductive models. Both physics-based and data
driven techniques can result in linear/nonlinear, static/dynamic,
and explicit or implicit models. Physics based techniques generally
result in continuous and deterministic models whereas; data-
driven techniques generally result in discrete and deterministic or
stochastic models.
While developing the physics based HVAC system models, dy-
namic models are commonly utilized for the slow moving tem-
perature and humidity processes (e.g., zone temperature dynamics,
zone humidity dynamics, and tank water temperature dynamics),
and static models are utilized for the fast moving dynamics [e.g.,
mixed air temperature and carbon dioxide (CO2) concentration in
mixing box, and ow rate of air and water through damper and
valve respectively] and energy consumption (fan or pump energy
consumption). Both static and dynamic models can also be devel-
oped for the same subsystem, e.g., the dynamic model of cooling/
heating coil can be developed by energy balance on the water and
air ow streams resulting in a pair of differential equations, alter-
natively, the heating/cooling coil static model can be developed by
considering the coil as a constant effectiveness heat exchanger.
Generally, physics based dynamic models are developed by
thermal-network method. In this method, the heat transfer in
HVAC components is often modeled by an electrical network in
which resistors and capacitor represent thermal resistivity and
capacitance, respectively, while current and voltage represent heat
transfer and temperature, respectively.
Data driven models are divided into the approaches based on
frequency domain models with dead time, data mining algorithms,
FL, stochastic models, statistical models, state-space models, case-
based reasoning models, geometric models, stochastic models
and instantaneous models. The grey box approaches differ based on
the selection of a particular parameter estimation algorithm; most
common of which are nonlinear least squares, simplex search, and
genetic algorithm (GA).
Table 1 shows the list of symbols and Table 2 provides the
comparison of the models based on the criteria of auto tuning,
robustness to parameters and disturbances, and ability to model
linear or nonlinear, single-input single-output (SISO) or multiple-
input multiple-output (MIMO) systems. In the following sections,
each class of approaches will be discussed in more details.
2.1. Data driven models
Major techniques used for data driven modeling of HVAC sys-
tems are shown in Fig. 1. These techniques are not limited to the
modeling of HVAC systems and can also be applied for the
modeling of other systems whose input and output measured data
is available. The following subsection discusses the data driven or
black box models in detail.
2.1.1. Frequency domain models with dead time
Many of the processes in the HVAC system (e.g., dynamics of
zone temperature and humidity) are slow moving with time delay
due to the heavy thermal inertia of the system. Such processes can
be modeled using the rst and second order (over-damped) models
with dead time [20e23], e.g.,
Ys K
Gs eLs (1)
Us ss 1
Ys 1
Gs 2 eLs (2)
Us as bs c
The symbols are dened in Table 1. The rst and second order
models are developed for SISO systems and can be extended to
MIMO systems. Examples of the models developed in the literature
include rst-order process models for AHU in Ref. [20], cooling coil
dynamics, and duct transfer function in Ref. [24], temperature
sensor in Ref. [21], ow meter in Ref. [22], and temperature control
dynamics of AHU in Ref. [23]. To identify the model parameters
with dead time, the process identication techniques are applied,
e.g., for models of off-coil air temperature, duct static pressure,
room temperature, and room pressure in Ref. [20].
First and second order frequency domain models with dead
time have simple structure and very few parameters to be deter-
mined from the measured data. Due to the wealth of the literature
on rst and second-order systems, the controller design is also
straight forward. On the other hand, the data gathering process is
intrusive and requires the interruption of the normal operation of
the system. Such models are applied to linear and time-invariant
(LTI) systems and thus complicated and time varying system dy-
namics cannot be represented with such models.
2.1.2. Data mining algorithms
A wide variety of data mining and machine learning algorithms
such as ANN and support vector machine (SVM) have been applied
to the HVAC systems modeling. This class of approaches is often
used for modeling complicated and nonlinear system dynamics.
 A. Afram, F. Janabi-Shari / Applied Thermal Engineering 67 (2014) 507e519 509

Table 1 Table 1 (continued )

List of symbols.
comp Compressor
Alphabets cond Condenser
adp, bdp, cdp, fdp Damper coefcients conv Convective
a1, b1 Coefcients that apply over a certain range of zone d Water below thermocline
air temperature dp Damper
A Area ele Electrical
cov Covariance evap Evaporator
C Thermal capacitance hp Heat pump
Cpa Specic heat of air at constant pressure i ith surface
Cpw Specic heat of water at constant pressure int Internal
COP Coefcient of performance iw Inlet water
d Polytropic index IA Inltration air
D Diameter j Time step
E Energy used max Maximum
f Fraction of the fan/pump power converted to the mec Mechanical
uid thermal energy mw Make-up water
h Convection coefcient o Outside air
H_ Enthalpy ow rate ow Outlet water
K Static gain ref Refrigerant
L Apparent dead time of the process rpm Revolutions per minute
m Mass ow rate sd Supply duct
M Mass sdi Inside the supply duct
n Number of observations sdo Outside the supply duct
N Valve authority surf Surface
Ncell Number of operating tower cells sw Fully-mixed sump water
P Pressure S Solar
q1 Back shift operator SA Supply air
r Radius u Water above thermocline
R Thermal resistance v Valve
Q Heat transfer rate vol Volumetric
Qevap Duty of evaporator w Water
Qleak,evap Energy gain in evaporator wai ith wall
Qleak,comp Energy loss in compressor wdi ith window
T Temperature wm Water and metal
u(t) Input z Zone
~ t
u Valve opening
U Conduction heat transfer coefcient
Uamb Conduction heat transfer coefcient to ambient
Uud Conduction heat transfer coefcient between
The network is trained by a supervised learning (often back-
upper and lower water layers propagation) algorithm. A SVM-based approach projects the non-
UA Overall conduction heat transfer coefcient linearly separable data into higher dimensional feature space
U_ Internal energy of the water through a mapping function where it can be separated linearly. A
V Volume
comprehensive survey of ANN applications in building energy
Vr Volume ratio of compressor
Vs Specic volume systems was provided in Ref. [25].
Vst Saturated volume of the compressor A single ANN or multilayer perceptron (MLP) can produce
w(t) Noise erroneous results due to non-optimum weights obtained during
W Power training. To avoid this situation, MLP ensemble was used in which
y(t) Output
y Measured value
multiple MLPs were used and the decision was made based on the
y* Predicted value majority vote. HVAC energy consumption model was developed in
y Average value Ref. [26]. Five data mining algorithms [i.e., exhaustive general chi-
z Height of the water square automatic interaction detector (CHAID), boosting tree,
Greek letters
random forest, SVM, and MLP ensemble] were used for the con-
a, b, g, d, q Polynomials
Effectiveness of heat exchanger struction of the model. MLP ensemble was found to outperform the
DP Change in pressure other algorithms.
DT Change in temperature Example of ANN in HVAC modeling include the one to estimate
DST Total internal entropy production the ground source heat pump (GSHP) well temperature in Ref. [27]
h Efciency
using the ambient temperature, inlet water temperature, time of
r Density
s Standard deviation operation, water ow rate, and exchanger heat capacity. ANN
s Time constant models for predicting the energy consumption of AHU and indoor
4 Angular position air quality (IAQ) sensors were developed in Ref. [28]. IAQ was
u Angular velocity
represented by indoor air temperature, humidity, and CO2 con-
Subscripts
a Air centration. The clustering of the ANN input variables produced a
ai Inlet air model with improved prediction accuracy as compared to the
amb Ambient model without clustering. The model of HVAC system based on
ao Outlet air dynamic ANN called nonlinear autoregressive exogenous (NNARX)
cc Cooling coil
was developed in Ref. [29]. ANN was used to model the cooling coil
cd Condenser
chw, r Chilled water returning to the chiller in Ref. [30] for application in fault detection. In this application, the
chw, s Chilled water supplied by chiller difference between the output of real coil and ANN was used to
cmp Campus detect the faults in the cooling coil. When there was no fault, the
cmp, r Campus water return
difference was low and when a fault occurred, the difference
cmp, s Campus water supply
became large.
510 A. Afram, F. Janabi-Shari / Applied Thermal Engineering 67 (2014) 507e519

Table 2
Comparison of different modeling techniques in building HVAC eld.

Modeling technique Criteria

Ease of Auto Robustness to Robustness to Model Prediction Generalization Linear/non-linear Tuning SISO/MIMO
tuning tuning parameters disturbances noisy accuracy capability model parameter models
data count

Frequency domain models M Y Y N N M L Lin L MIMO

with dead time
Data mining algorithms M N N N Y H L NL H MIMO
Fuzzy logic models M N Y Y N H M NL M MIMO
Statistical models H Y N N Y L M Lin H SISO
State-space models L N N N N H L Lin H MIMO
Geometric models M N N N N M M NL M SISO
Case based reasoning M Y Y N Y M M NL M MIMO
Stochastic model L N N N Y H L NL L MIMO
Instantaneous models M N N N N M L NL L MIMO
Physics based models L N N N N L H Lin/NL L MIMO
Grey box models L Y Y Y Y H M Lin/NL L MIMO

Note: Y e yes, N e no, H e high, M e medium, L e low, Lin e linear, NL e nonlinear.

Examples of SVM in building load forecasting include the esti-
mate of cooling load of the building HVAC system based on the
single variable time series [31]. The measured data on the cooling
load over a period was used to predict the future cooling load using
SVM. The global optima of SVM penalty parameter, intensive loss
function and kernel function were found using the ant colony
optimization (ACO).
The prediction capability of ANN and SVM approaches can be
improved by hybrid data mining algorithms. For example hybrid
ACO-SVM model outperformed the simple ANN model in Ref. [31]. A
hybrid SVM based autoregressive integrated moving average
(ARIMA) model was used for the building cooling load prediction in
Ref. [32]. The accurate prediction of the cooling load is mandatory
for the optimal control of the HVAC system [32]. The ARIMA model is
suitable for linear prediction and SVM is suitable for the nonlinear
prediction of the cooling load. The building cooling load was pre-
dicted in Ref. [33] using the kernel principal component analysis
(KPCA) and SVM. The input data was applied to the KPCA, which
extracts the nonlinear features from the input data and produces the
output data suitable for the SVM. The resultant KPCA-SVM model
had good generalization ability and low dependency on the training
data. Another method for cooling load forecasting is used in Ref. [34]
which was based on simulated annealing based particle swarm
optimization (SAPSO) algorithm to compute the parameters of the
SVM. The SAPSO-SVM resisted the premature convergence and
outperformed the simple SVM-based load forecasting.
Data mining algorithms are easy to implementation. They can
model noisy data with many variables and complex nonlinear re-
lations between inputs and outputs. Comprehensive literature is
available on parameter selection and training methods. Unlike the
ANN, the SVM is resistant to over-tting the data and has better
performance than simple ANN. SVM nds the global optimum so-
lution in the data and provides the best t for the data. In order to
build the models using data mining algorithms, large amount of
training and testing data is needed. No physical interpretation of
the developed model is possible and the performance degrades
when conditions deviate from training and testing conditions.
These algorithms are suitable for ofine model development and
online implementation of these algorithms is often cumbersome.
2.1.3. Fuzzy logic models
FL uses the expert knowledge in the model development. FL
models are developed by implementing the if-then-else state-
ments. The rules are written in the form of a table or database [35].
TakagieSugeno (TS) FL modeling method uses the local linear
models blended together to model the nonlinear system. Variations
of FL technique include fuzzy adaptive network (FAN) and adaptive
network based fuzzy inference system (ANFIS). FAN is a ve layered
network which utilizes the same back-propagation algorithm as
ANN for training of the network. ANFIS comprises of a multilayer
structure similar to ANN but unlike ANN each layer performs
different function. Its structure is more complicated than ANN or FL
alone and is a combination of both. The node architecture and
properties remain the same in a single layer and vary across the
layers. A review of FL modeling and control of HVAC systems was
presented in Ref. [16].
Examples of FL include FAN used to model the thermal comfort
in Ref. [36], TeS FL model of AHU in Ref. [37], and model of thermal
comfort in Ref. [38]. ANFIS was used to predict the fan speed and
damper opening using the zone temperature and ambient tem-
perature in Ref. [39]. The energy use of the two buildings was
predicted by Ref. [40] using the ANN and GA based ANFIS (GA-
ANFIS). In many HVAC systems, GA-ANFIS performs better than
ANN as it has more complex structure than the ANN. For large
datasets, the training time for GA-ANFIS is also greater than ANN
but is comparable at small datasets.
Models developed with FL are generally very simple and easy to
understand. FL incorporates the operators experience in model
design. This requires comprehensive knowledge of plant and its
various operating states. Alternatively, FL model development re-
quires large amount of performance data for training TeS FL
models, FAN and ANFIS. The required knowledge and data for FL
model development may not be readily available for many HVAC
components and thus presents a difculty in modeling these
components using FL.
2.1.4. Statistical models
The statistical black box models consist of single and multivar-
iate regression, autoregressive exogenous (ARX), autoregressive
moving average exogenous (ARMAX), ARIMA, nite impulse
response (FIR), Box Jenkins (BJ), and output error (OE) models. The
mathematical expression for the generalized structure of statistical
black box models in a simple input/output relationship is given
below [41].
Appropriate choice of numerator and denominator polynomials
of input u(t), output y(t) and noise w(t) results in different models
such as ARX, ARMAX, ARIMA, FIR, BJ and OE.

    
b q1 g q1
a q1 yt  1  ut  1  wt: (3)
q q d q
 A. Afram, F. Janabi-Shari / Applied Thermal Engineering 67 (2014) 507e519 511

Fig. 1. Data driven modeling techniques.

Some of these models such as FIR, OE and BJ do not consider the
output while the others such as AR and ARMA do not consider the
input in the input/output equation therefore may not be useful for
closed loop control system. The models ARX and ARMAX consider
both the input and output in their structure and are useful for the
design of the closed loop control system. The model ARMAX is
superior to ARX as it incorporates the time series of error in the
model structure which is essential for capturing the dynamics of
the error and better control performance. ARIMA is a generalization
of ARMAX, modeling the stationary and non-stationary data into a
single step, and consists of autoregressive, integrated and moving
average parts.
Examples of statistical models include single variable linear
model to calculate the energy demand of a variable capacity air
source heat pump (ASHP) and a GSHP based on the outside tem-
perature in Ref. [42]. The relation between AHU fan speed and
power consumption in Ref. [43] was modeled using a second order
polynomial whose coefcients are derived by curve tting on the
measured data. A more accurate model can include multiple input
variables. The model of room temperature and humidity based on
the time series regression of the internal and external variables
such as ambient temperature, solar irradiance, electricity con-
sumption, zone temperature and zone humidity was presented in
Ref. [44]. The developed model can predict the room temperature
512 A. Afram, F. Janabi-Shari / Applied Thermal Engineering 67 (2014) 507e519
very accurately many days ahead under similar conditions of
training data. The ARX model of the energy use and zone temper-
ature (as a function of the temperature set point and ambient
temperature) were developed in Ref. [45]. The autoregressive in-
tegrated (ARI) model for the compressor capacity and power con-
sumption was developed in Ref. [46]. The ARX model, ARIMA
model, Bayesian networks and ANN for load forecasting in air-
conditioned non-residential buildings were compared in Ref. [47]
concluding that the ARX model outperformed the other three
models.
Statistical methods such as linear and polynomial time series
regression provide steady state modeling which is useful for quick
calculations. The model can be obtained with limited data. These
models have low accuracy and are unable to capture the nonlinear
dynamics of the system. Since the processes in an HVAC system
depend on their previous values, a time series regression model
(i.e., ARX, ARMAX, and ARIMA) captures these correlations by
including the process variables from the previous sampling times.
This results in a very accurate model of the process dynamics but
the memory requirement increase in order to save the previous
values. Correlation analysis must be performed in order to deter-
mine the number of previous samples that are important and need
to be used in the time series regression model of a process. These
statistical methods also work well under time varying process
noise. Many useful online tuning algorithms exist for these
methods. MIMO model identication using these methods requires
large number of parameters to be determined. Model tuning re-
quires experience as brute force tuning takes a long time to tune the
model parameters.
2.1.5. State-space models
This class of black box models derives the system model from
data directly into the state-space form. The sub-space state-space
(4SID) model determines the sequence of the states and system
matrices from the input and output data measurements. It requires
only one parameter, i.e., the number of block rows of Henkel matrix
to be supplied by the user for model tuning [48].
A 4SID model is a deterministic technique and does not use the
recursive algorithms for tuning hence requiring short time to tune
the model. A 4SID model complexity does not increase and can be
easily extended to MIMO systems. However, a 4SID approach needs
lot of input and output measurement data for tuning the model so
online implementation is complicated. The model developed with
4SID technique performs poorly if the process noise has time
changing properties.
2.1.6. Geometric models
Geometric modeling methods deal with the construction and
representation of the curves, surfaces and volumes. Thin plate
spline (TPS) is part of the geometric modeling methods. The name
of TPS comes from the shape of the function that resembles a bent
thin plate with smooth surface. TPS has the smooth features and
good ability of extrapolation. TPS maps the input output data using
a function which contains the weighted sum of Greens function
and linear regression terms [49]. TPS compared to ANN has its own
advantage as it is independent of initial conditions, however, unlike
the ANN, it is sensitive to the noise present in the data on which it
interpolates. It also has the disadvantage of being computationally
expensive as the data size increases [49].
2.1.7. Case-based reasoning
Case-based reasoning (CBR) models make use of the similar past
cases to construct the models [50]. Topological case-based
modeling (TCBM) is a special CBR technique in which the models
are built on a subset of data instead of the whole data. The
algorithm of TCBM nds the cases whose inputs are the most
similar to the new inputs and averages the outputs of the similar
cases as the corresponding new output [49]. Once the real outputs
are obtained, it adds them to the case-base and adapts to the new
situation. If a previous case is not found in the database, the output
is computed based on the neighborhood cases.
CBR approach is suitable for a nonlinear and ill-dene system. It
reduces measurement noise and compresses the data. The models
mature with time and add data to the case-base online to increase
the future prediction performance. The case-base should be
comprehensive to cover all the input cases. CBR technique suffers
from the problems related to the unseen cases.
2.1.8. Stochastic models
Stochastic models deal with the random processes. Some pro-
cesses in the HVAC systems act as random variables and can be
modeled using the probability density functions (PDF). The energy
used by the fans of a large hypermarket based on the operating
schedule was modeled in Ref. [51] using the Gaussian PDF. The
variable speed fans were assumed to be operating at the minimum
power level during the unoccupied hour and the energy con-
sumption during the occupied hours was assumed to follow a
normal distribution function.
Many physical processes can be approximated to standard
normal and uniform distributions. However, a large amount of data
is required to obtain the accurate shape of the PDF of a random
variable. The model predictions suffer if the PDF is not modeled
properly.
2.1.9. Instantaneous models
Just in Time (JIT) model comprises of a statistical model and a
pattern model. JIT approach uses previous data to nd the patterns
similar to the current data. It combines the statistical modeler with
the pattern recognition modeler and provides good prediction re-
sults. For example, in Ref. [52] the load pattern model algorithm
searches the database for the similar weather condition to todays
and peak load statistical model with todays weather prediction
adjusts todays load pattern. JIT model can be applied for modeling
the other HVAC components such as cooling coil model, and chiller
model from the measured data. In order to implement JIT technique
successfully, large database is needed. Missing data creates major
problem and if the number of parameters is large, the search could
take longer.
2.2. Physics based models
Physics based models are also known as analytical rst principal
models, forward models or white box models. These models are
based on the detailed knowledge of the process and its underlying
physical principles. They require signicant effort to develop and
calibrate. Though the physics based models usually take the form of
time-domain differential equations, they can be readily converted
to frequency domain transfer functions [53] or time-domain state-
space representations [54,55]. Transfer functions and state-space
models developed from the physical equations are different from
the data driven modeling approaches discussed above which do not
have underlying physical meanings. Major applications of physics
based modeling approach to HVAC components are discussed in
the following subsections.
2.2.1. Zone model
The zone temperature is maintained by adding or removing the
heat to balance the internal and external gains and losses [38]. The
zone model can be obtained by the energy balance of a room in the
steady state [55]. Heat is transferred to the zone through the supply
 A. Afram, F. Janabi-Shari / Applied Thermal Engineering 67 (2014) 507e519
air, conduction through walls and windows, inltration and supply
air, and internal and external gains due to the humans and solar
ux [56]. Heat transfer to a zone is commonly modeled by using
heat conduction equation model, heat balance method, weighting
factor (response factor or transfer function method), and thermal-
network method [57,58].
In heat balance method [58], the energy ow is modeled using
the rst law of thermodynamics (i.e., conservation of energy). For a
zone, generally a heat balance equation is written for each heat
transfer element (wall, window, ceiling, oor etc.) and zone air.
These equations are simultaneously solved to nd the unknown
temperature of the zone and surface of each heat transfer element
using the matrix algebra techniques. Once the surface temperatures
are determined, convective heat transfer to and from the zone air
can be calculated. Zone temperature can be calculated by the
following equation at a time step j
PN
a1 i1 Asurf;i hconv;surf;i Tsurf;ii;j ra Cp VIAj Toj ra Cp VSAj TSAj Qconv;intj
Tzj P (4)
b1 N i 1 Asurf;i hconv;surf;i ra Cp VIAj ra Cp VSAj
The coefcients a1 and b1 apply over a certain range of zone
temperature. Similar equations need to be written for each element
enclosing the zone and solved simultaneously to compute the zone
air temperature. Once these coefcients are determined, the heat
transfer rate to the zone can be calculated as follows
Qzj a1 b1 Tzj
The heat balance method can be expanded to multiple zones but
the size of the coefcients required for solving the simultaneous
equations becomes too large resulting in increased computation
time.
In weighting factor method, heat gains and losses are calculated
using the z-transfer functions. Two types of weighing factors i.e.
heat gain weighting factors (relating zone cooling load to instan-
taneous heat gain) and air temperature weighting factors (relating
zone air temperature to the total load of the zone) represent the
transfer functions. The cooling load at time j is calculated as
Qj v0 qj v1 qj1 .  w1 Qj1  w2 Qj2 .:
The cooling load at time j is the linear combination of present
and past values of instantaneous heat gain (qj, qj1,...) and previous
values of cooling load (Qj1, Qj2,...). The weights of the linear
combination (v0, v1,..., w1, w2,...) are the heat gain weighting factors.
Heat gain weighting factors are different for each heat source due to
the variation in amount of convective and radiation energy leaving
each source. Heat gain weighting factors also differ across buildings
since there is a variation in thermal properties of various materials
used during construction. Zone air temperature at time j can be
calculated as
       In an HVAC system, the heating/cooling coil conditions the
Tzj 1 g0 Qj  ERj P1 Qj1  ERj1 P2 Qj2  ERj2

.  g1 Tzj1  g2 Tzj2 :
Zone temperature at time j is the linear combination of present
and past values of cooling load (Qj, Qj1,...), energy removal rate (ERj,
ERj1,...) and deviation of air temperature from set point
513
Tzj1 ; Tzj2 ; .: The weights of linear combination (P1, P2,..., g0, g1,...)
are called temperature weighting factors and typical values for
light, medium and heavy room envelope construction are provided
in Ref. [58].
While using the weighting factor method, it is assumed that the
heat transfer processes are linear and the system properties inu-
encing the weighting factors and not changing with time. This
could be limiting the application of this approach.
In thermal-network model, the building is divided into a
network of nodes with interconnecting paths through which the
energy ows. The implementation of this method varies based on
the selection of nodes on which energy balance is applied. This
method is considered as the renement of heat balance method.
Temperature for a single zone using the lumped parameter model is
given in Ref. [24] by applying the heat balance on the zone air as
follows
dTz
ra Vz Cz mSA Cpa TSA  Tz Uwai Awai Twai  Tz
dt
 
Uwdi Awdi Twdi  Tz QS Qint : (8)
Following are some of the examples of various implementations
of thermal-network model. Models of building envelope can be
developed by applying the heat balance on the outside surface and
(5) inside surface of the zone in conjunction with the zone air [43]. In
Ref. [59], room was modeled by applying the energy balance on
walls, two layers of oor and zone which yielded four equations.
Two layers of oor help to model the radiant oor heating system.
The zone temperature model in Ref. [53] was obtained by energy
balance on the zone, two walls, and the ceiling also resulting in four
equations. The models discussed above only consider the heat
transfer to the zone through the process of convection (by venti-
lation, air inltration and exltration) and conduction (through
walls, windows, oors and ceiling). The radiant exchange between
surfaces or neighboring zones occurs if they are at different tem-
peratures. The radiant exchange between two surfaces is provided
(6) by the StefaneBoltzmann law and becomes signicant when the
surfaces are at signicantly different temperatures or the energy
transfer to a zone is calculated for a longer period of time (e.g., a
year) and cannot be ignored [60].
Humidity ratio and CO2 concentration in a zone was modeled in
Refs. [53,61]. These models take into consideration the CO2 and
humidity generated by humans. The rate of change of CO2 and
humidity in a zone is dependent on the supply air ow rate and
difference between the zone and supply air concentration as well as
the number of persons present in the zone.
2.2.2. Cooling and heating coil model
supply air to maintain the indoor temperature. Reheat coil is used
in the variable air volume (VAV) box to control the temperature of
the supply air. The cooling coil dynamic temperature model can be
(7)
obtained by the energy balance on the water and air side of the coil
[62e64,53]. Mass balance on the air side gives the humidity ratio of
the outlet air. The temperature of outlet water and air is given as
follows
514 A. Afram, F. Janabi-Shari / Applied Thermal Engineering 67 (2014) 507e519

dTow coefcient, pressure difference across the damper and the position
Cwm mw Cpw Tiw  Tow  UAcc Tow  Tao (9) dependent ow cross section area of the damper.
dt
q
dTao ma cdp ra DPdp Adp 4 (17)
Cpa UAcc Tow  Tao  ma Cpa Tao  Tai : (10)
dt
In Ref. [65] heating system was modeled by calculating the heat
2.2.6. Valve model
delivery rate of the radiators due to convection and radiation. The
Valves control the ow rate of water and refrigerant in HVAC
cooling/heating coil can also be modeled as a constant effectiveness
system. Expansion valve is used in chiller to control the ow of the
heat exchanger [6,66,67]. This type of steady state model is very
refrigerant from condenser to evaporator. Valve is also used to
simple and determines the temperature of the outgoing air and
control the ow rate of the water to cooling or heating coil. The
water from the inlet air and water temperature as follows
nonlinear hydraulics of the plant are modeled using the valve travel
Qmax vs. the water ow rate graph in Ref. [21]. A linear model of the valve
Tow Tiw  (11) relating the control signal to the valve opening is provided in Ref.
mw Cpw
[63]. The valve opening u ~ t depends upon the control signal
u(t) [0, 1] and the valve authority. The valve opening and water
Qmax ow rate is given as follows
Tao Tai  : (12)
ma Cpa
ut
~ t p
u (18)
where Qmax maCpa(Tai  Tiw) when maCpa < mwCpw, otherwise u2 t1  Nv Nv
Qmax mwCpw(Tai  Tiw).
~ tmw;max :
mw u (19)
2.2.3. Duct and pipe model
The conditioned air is supplied to the zone and exhaust air is The model of a temperature regulating valve for hydronic radi-
extracted from the zone through the ducts. The air passing through ator is given in Ref. [59]. The relation between the valve opening and
the ducts loses its energy to the surrounding environment due to the water ow rate can be modeled by a third order polynomial
the heat transfer loss through the walls of the duct. The supply duct equation with constant coefcients. The refrigerant ow rate in a
model represents the heat transfer between the air inside the duct chiller expansion valve is modeled using orice equation in Ref. [68]
and the ambient [55] given as follows as follows
  n
dTsdi 4Usd Tamb  Tsdi mref cvalve rcond Pcond  Pevap : (20)
: (13)
dt Csd Dsd ra
The curve of ow coefcient cvalve is provided by the manufac-
The supply air duct model in Ref. [53] gives the variations of the turer and n is a constant.
outlet temperature with respect to the inlet temperature of the air
as follows
2.2.7. Fan and pump model
Fans control the ow rate of air and pumps control the ow rate
dTao Usdi Usdo ma Cpa
Tai  Tao : (14) of water and refrigerant through the HVAC system. For example, in
dt Usdi Msd Csd
GSHP, refrigerant is circulated in the ground loop and in AHU, water
is circulated through cooling/heating coils using pumps. The power
consumption of the fan or pump depends on the ow rate, pressure
2.2.4. Mixing box model difference between inlet and outlet and efciency of the fan or
A portion of the return air from the zone is mixed with the fresh pump [69,22,70,71].
air from outside in a mixing box and mixed air is supplied to the The temperature of air passing through fan increases due to the
zone after passing through heating/cooling coil [62]. The mixed air inefciencies of the fan motor if the motor is installed in the air stream.
temperature is the linear combination of return air and fresh The temperature of the air at the outlet of the fan is given by Ref. [6]
outside air temperature [53] given as
Wfan f
Tao Tai : (21)
mo To mSA  mo Tz ma Cpa
Tm : (15)
mSA
By replacing all temperature variables with the corresponding 2.2.8. Storage tank model
CO2 concentration or humidity ratio variables in the above equa- Storage tank stores the chilled and hot water to be supplied to
tion, one can get the similar linear relation for nding the mixed air the cooling and heating coil [72]. The temperature at the outlet of
CO2 concentration [61] or humidity ratio [66], respectively. the tank is given as the summation of the heat transferred by the
inlet water, heat supplied by the heat pump and heat lost to the
2.2.5. Damper model ambient as follows
The damper controls the ow rate of the air in HVAC system. The
dTow 1 h
mass ow rate of the air passing through the damper depends on
 mw Cpw Tiw  Tow Ehp Ehp;max COP
the control signal u(t) [0, 1] which controls the damper opening dt rw Cpw Vtank
[22]. i
UAtank Tamb  Tiw :
ma adp ebdp ut fdp (16) (22)
In the damper model presented in Ref. [39], the mass ow rate A more comprehensive model dealing with the charging and
of the air through the damper depends upon the damper ow discharging modes of a stratied thermal storage tank was
 A. Afram, F. Janabi-Shari / Applied Thermal Engineering 67 (2014) 507e519 515

provided by Ref. [73]. A large university campus was supplied by a
30 m high thermal storage tank. The thermocline layer which
separates the hot water from the cold water, formed at about 20 m
height. The water was warm above the thermocline and cold below
it due to the return water from the campus entering the top of the
tank. The chiller supplied the water near the bottom of the tank and
the water supplied to the campus was also from the bottom of the
tank. During charging mode, the ow rate of the water supplied by
the chiller was higher than the ow rate of the water supplied to
the campus. During discharging mode, the ow rate of the water
supplied by the chiller was lower than the ow rate of the water
supplied to the campus. Despite the different ow rates of the in-
dependent charging and discharging circuits, the water volume in
the tank remained constant as the supply and return water ow
rate in both the charging and discharging circuits was equal. The
tank dynamics during the charging mode are given as
be modeled separately. The chiller works on the basis of vapor
compression cycle (VCC) [68]. In VCC refrigerant gains heat from
the conditioned zone and gets evaporated in the evaporator. The
compressor increases the pressure of refrigerant making it super-
heated and releases it into the condenser where the heat is rejected
to the water or air. Then the expansion valve reduces the pressure
by releasing the refrigerant in the evaporator in a cool state and the
cycle continues.
Both condenser and evaporator can be modeled as classical heat
exchangers [75]. The model of the compressor outlet pressure and
mass ow rate were given by Ref. [76] as follows
" #
Wcomp d

Pcond Pevap Vrd Vr hmec hele  Pevap Vrd1  1
mref Vs d  1
(29)

H_ d md Cpw Tchw;s mref Vst rref hvol urpm : (30)

H_ u mu Cpw Tu The chiller power consumption is dependent on the water ow

(23) rate, heat capacity of water, temperature difference between
Tcmp;s Tchw;s entering and leaving water in chilled water loop and COP of the
Tcmp;r mcmp;r  Tu mu chiller [69]. COP of the chiller varies based on the load on the chiller.
Tchw;r : Ref. [61] provided the model for computation of COP from PLR. In
mchw;r
this model [61], COP was high when PLR was low and vice versa.
The tank dynamics during the discharging mode are given as Ref. [77] compared the performance of ASHRAE primary toolkit
model, Gordon-Ng universal chiller model and DOE-2 model. First
H_ d md Cpw Td two chiller models are physics based whereas the last model is
empirical model. The two physics based models differ in the
H_ u mu Cpw Tcmp;r
methodology and assumptions made during the development of
Tchw;s mchw;s  Td md (24)
the models. All three models were found to be equally accurate.
Tcmp;s
mcmp Gordon-Ng universal chiller model provides the linear performance
equation as follows
Tchw;r Tcmp;r :

  
The internal energy of the tank is Tevap;in 1 Tevap;in T T
1 1 DST Qleak;eqv cond;in evap;in
Tcond;in COP Qevap Tcond;in Qevap
U_ u H_ u Q_ d>u Q_ amb>u

(25)
U_ d Hd Q_ u>d Q_ amb>d :
RQevap
1
1
Tcond;in COP
where Q_ amb>u and Q_ amb>d are the heat transfer rate in Watt from
1 1
the ambient to the warmer and cooler water in the tank respec- R
tively given as cond mcond Cpw evap mevap Cpw

Qleak;comp Tevap;in
Q_ amb>u Tamb  Tu 2prtank zu Uamb Qleak;eqv Qleak;evap :
(26) Tcond;in  Tevap;in
Q_ amb>d Tamb  Td 2prtank zd Uamb
(31)
and Q_ u>d denotes the heat conducted from warmer water to cooler
water and Q_ d>u represents the heat conducted from cooler water to
warmer water in the tank given as 2.2.11. Heat pump model
Heat pump provides energy from a source of heat to a sink of


heat. Most common categories of heat pump are air source heat
Q_ u>d Tu  Td prtank
2 Uud

 (27) pump (ASHP), water source heat pump (WSHP) and ground source
_
Q d>u Td  Tu prtank Uud :
2
heat pump (GSHP). In the cooling mode, heat pump extracts heat
from a conditioned zone or thermal storage, and rejects it to the
ambient air, water or ground depending on the type of heat pump.
2.2.9. Boiler model
In the heating mode, the operation of the heat pump reverses and
Boiler consumes fuel and produces hot water. The temperature
the heat is extracted from the ambient air, water or ground and
of the water in a boiler is given by the heat supplied and the dif-
rejected to the conditioned zone or thermal storage. Heat pump
ference in supply and return water temperature [74]
also works on the basis of vapor compression cycle as described in
dTboiler 1   the above section. As described above, evaporator and condenser
Q mw Cpw Tiw  Tow : (28) can be modeled as classical heat exchangers. The models of
dt mw Cpw heat
expansion valves and compressor are provided in the Section 2.2.6
and Section 2.2.10, respectively.
2.2.10. Chiller model Analytical models of ground heat exchanger (GHE) of a GSHP
The chiller can be broken down into its individual components based on line source theory and cylindrical source theory were
evaporator, condenser, compressor and expansion valve which can reviewed in Ref. [78]. The later model is more accurate than the
516 A. Afram, F. Janabi-Shari / Applied Thermal Engineering 67 (2014) 507e519
former model due to its distinct physical meanings and forms the
basis of many numerical simulation models. Above mentioned
analytical models along with numerical models of GHE (e.g., nite
difference method, nite element method, nite volume method,
response factors model, and load aggregation algorithms) were also
reviewed in Ref. [79]. A comprehensive review of WSHP models
(i.e., analytical models, empirical models and models falling be-
tween the two extremes) was provided in Ref. [80].
2.2.12. Cooling tower model
Cooling tower is used to reject heat from the chiller condenser.
Fans and pumps in the cooling tower are controlled to maintain a
desired water temperature entering the condenser. Cooling tower
can be modeled as a heat exchanger in which the heat is rejected to
the air through natural or forced convection. Physical models of
cooling tower are developed using Merkels method and
Effectiveness-NTU method, whereas empirical model is developed
by Stoeckers method [81].
The heat rejection of the tower cell and sump water tempera-
ture based on the Effectiveness-NTU method is given by Ref. [82].
Heat rejection is given as
Qtower a ma hao  hai : (32)
Water is lost in the operating tower cells due to evaporation. To
keep the level of water in sump reservoir constant, water enters the
sump reservoir from water make-up source. The ow of the water
make-up is equal to the total water loss from the cells. The sump water
is considered to be fully mixed and its temperature is given as [82]
" !
dT X N
cell X
Ncell
rw sw mow Tow  Tsw k miw  mow k
dt
k1 k1
# phenomenon could be very complicated to model and for other
Tmw  Tsw : (33)
Physics based models provide good generalization capability but
lack the accuracy compared to the data driven models. The cali-
bration of the physics based models also presents a difcult chal-
lenge due to the identication of large number of parameters.
2.3. Grey box models
Grey box models use physics based models as the model
structure whose parameters are estimated from the measured data.
were estimated from the data obtained by a comprehensive zone
model developed in EnergyPlus.
The physics-based ARMAX model of room temperature was
developed in Ref. [56]. The model was able to predict the room
temperature quite accurately up to ten weeks ahead. A state space
model of the room with zone temperature and outside air tem-
perature as the state variables and HVAC power as the input was
given in Ref. [86]. The system matrices were obtained by applying
least squares method on the measured data.
The continuous model of the cooling coil and chilled water ow
rate control valve was developed in Ref. [63]. The grey box models
of the electronic expansion valve, thermostatic expansion valve,
compressor and evaporator were developed in Ref. [87]. The
parameter estimation of the grey box models was performed by the
numerical search methods nonlinear least squares and simplex
search. The parameters were successfully identied from the
measured data using both methods. The cooling coil model was
given in Refs. [88,89] whose parameters can be determined by
manufacturers data or by real time experimental data. The cooling
coil model for water and air temperature as well as the coil capacity
whose parameters can be determined from the real time experi-
mental data is given in Ref. [82]. The model for power consumption
of the chiller as a function of chilled water temperature was given in
Ref. [90]. The power of the chiller was a function of water tem-
perature change and voltage. The model parameters were found
using the least squares estimates on the measured data.
Grey box models provide good accuracy than physics based
models and better generalization capability compared to data
driven models but they are also the hardest to develop. In order to
develop grey box models, both the knowledge of underlying
physical phenomenon and inputeoutput data of the system is
required. For some HVAC subsystems, the underlying physical
systems, the inputeoutput data may not be readily available mak-
ing it difcult to develop the grey box models. The parameters for
grey box models also need retuning when the operating conditions
deviate from the training data in order to ensure higher accuracy.
3. Performance comparison metrics
In order to compare the prediction results of different models
and their deviation from the measured data performance metrics
are dened. Following metrics are used for comparing the perfor-
mance of different models [26,28,29,32,39,40,56,91e94].
Absolute error [26,92]
These models provide physical meaning and are useful for control  
 
applications with generalization capability superior to data driven AE y*  y (34)
models. Grey box models also capture the effects of any un-
modeled dynamics of the system which were left out of the Absolute percentage/relative error [26,29,92]
 * 
modeling equation by choosing appropriate parameters. Knowl- y  y
APE   (35)
edge of both the physical phenomena and the data from the process y 
is needed to nd the parameters.
The zone model in Ref. [83] was represented by a resistive Maximum absolute error [56,91]
capacitive (RC) circuit. Genetic algorithm was used to nd the  
 
model parameters such as thermal resistances and capacitances MAXAE maxi yi  y*i  (36)
with best t to the measured data. The lumped capacitance model
of the thermal zone in Ref. [84] was obtained using the heat balance Mean bias error [93]
equations on zone air and wall. The parameters of this grey box
model were identied from the indoor temperature, outdoor 1Xn
* 
MBE i1
yi  yi (37)
temperature and energy consumption. These parameters can be n
obtained using the construction data of the building and can also be Mean absolute error [26,28,56,91,92]
estimated from the measured data using a parameter identication
method. In Ref. [85], the zone model was represented by RC circuit 1Xn
MAE AE (38)
whose parameters were identied by sequential quadratic pro- n i1 i
gramming (SQP) optimization algorithm. The model parameters
 A. Afram, F. Janabi-Shari / Applied Thermal Engineering 67 (2014) 507e519 517

Mean squared error [56,92] the actual outcomes. Since average of a purely random error is zero
therefore, the values of MBE, MAE, MSE, RMSE, MAPE, and RME
1Xn
* 2 average errors close to zero indicate a more useful prediction.
MSE y i  yi (39) Compared to other mean errors dened above, MSE has advantages
n i1
such as it is analytically tractable and measures the precision
Root mean square error [39,93,94]
(variance) and accuracy (bias) [95].
p StdAE and StdAPE measure the dispersion (variance) of AE and
RMSE MSE (40) APE from their respective means. CV is the normalized measure of
Mean absolute percentage/relative error [26,28,29,92] the dispersion of the error probability distribution from its mean
value. Lower values of StdAE, StdAPE, and CV mean that the model
1Xn predictions are closer to the observations. The dispersion of error is
MAPE APEi (41) used to see the trend in error (i.e., increasing, decreasing or con-
n i1
stant). A purely random error exhibits a xed dispersion and thus
Standard deviation of absolute error [26,28,29,92]
has a constant distribution. Obviously, the dispersion of an
r increasing error also increases and that of a decreasing error also
1 Xn
StdAE AEi  MAE2 (42) decreases.
n  1 i1 CC measures the dependent relationship or strength of associ-
Standard deviation of absolute percentage error [26,28,29,92] ation between two datasets (i.e., model predictions and actual
measurements). If they are highly correlated, the value of CC will be
r
1 Xn close to 1 for positively correlated or 1 for negatively correlated.
StdAPE APEi  MAPE2 (43) As the relationship between the two datasets decreases, the value
n  1 i1
of CC starts to change and approaches to 0 for completely uncor-
Coefcient of variation (RMSE) [40,93,94] related datasets. R2 is the square of CC and ranges from 0 to 1. G, D,
r and R2 describe how well a model ts the measurements. While the
1 1Xn
* 2 RMSE values of G range from 0 to 100, the values of D and R2 range from
CV i1
yi  yi (44)
y n y 0 to 1. Higher values correspond to a better t of the model pre-
dictions to the measurements. Both D and R2 produce similar values
Correlation coefcient [91]
on the same data set and can be used alternatively.
  In addition to the above mentioned methods, ASHRAE standard
cov y; y*
CC (45) [96] provides the methods for model validation and testing for
sy sy* building energy performance simulation programs. In order to
Goodness of t [92] comply with this standard, the software has to successfully pass the
comparative tests, analytical verication, and empirical validation
0 r
2 1
Pn
* of building envelope and mechanical equipment simulations [97].
B i1 y i
 y i C Home energy rating system building energy simulation test (HERS
G B C
@1  q A  100 (46)
Pn  1
P n  2
BESTEST) [98] provides the guidelines for developing the pass/fail
i1 yi  n i1 yi criteria for energy simulation programs based on the comparison
with reference programs (i.e., BLAST, DOE2 and SUNCODE). ISO
Coefcient of determination [56] standard [99] provides the methods for calculation of room tem-
perature, with standard [100] outlining the general criteria and
 P P P * 2
n yi y*i  yi yi validation methods for room temperature calculations.
D

 

 (47)
P 2 P 2 P *2 
P * 2
n yi  yi n yi  yi 4. Conclusions

Coefcient of multiple determination [39]
Pn  2
i1 y*  yi
R2 1  Pn i 2
i1 yi
AE is the absolute difference between the value predicted by the
model and the measured value at each sampling instant. While raw
difference between predicted and measured value can take both
positive and negative values, AE is always positive. APE is the
normalized error and can be used to compare the model perfor-
mance across different datasets. Errors are also called residuals and
residual analysis is used in statistical regression to determine the
prediction ability of a certain model. If there is no pattern in the
residuals and they seem random, then the model is a suitable t. If
the residuals show a pattern (e.g., a U-shape, an inverted-U shape)
then a linear model is not suitable for the data and a more suitable
nonlinear model is desirable.
While MAXAE computes the maximum value of absolute error,
MBE, MAE, MSE, RMSE, MAPE, and RME compute the averages of
bias, absolute, squared, root-squared, absolute percentage, relative
and absolute relative errors and measure how close forecasts are to

In this paper, the general techniques used for modeling HVAC
systems are reviewed. Three main classes are identied and
include data driven, physics based and grey box methods.
(48)
The data driven methods completely rely on the measurement
data of the input and output variables and t the linear and
nonlinear functions to approximate the behavior of the system
as close as possible. These models include well established
techniques such as frequency domain models with dead time,
data mining algorithms (e.g., ANN and SVM), FL models (e.g.,
FAN and ANFIS), and statistical models (e.g., ARX, ARMAX and
ARIMA).

On the other hand physics based models completely rely on the
knowledge of the process and the physical laws governing this
process. The models built using physics based methods repre-
sent the system more closely and have better generalization
capabilities. The data driven models tend to degrade as the
conditions vary from the training data.

Another modeling technique known as the grey box models
exists at the intersection of the physics based and data driven
models. Grey box models use physical laws to dene the overall
structure of the models and use measured data to nd the
518 A. Afram, F. Janabi-Shari / Applied Thermal Engineering 67 (2014) 507e519
parameters of these models. For parameter identication, many
optimization techniques such as least squares, gradient descent
and GA are used.

In order to compare the performance of the models several
performance comparison metrics are used in the literature. Such
metrics have been collected from the literature survey and re-
ported in this paper.

The qualitative comparison of different modeling techniques is
also carried out based on their strengths, weaknesses, ease of
tuning, robustness and ability to model linear/nonlinear, and
SISO/MIMO systems.
Acknowledgements
This research was nancially supported by Ryerson Center for
Urban Energy (CUE), Toronto Hydro and Mitacs-Accelerate
Program.
References
[1] ASHRAE, Handbook-fundamentals, American Society of Heating, Refriger-
ating and Air-Conditioning Engineers, 2009.
[2] ASHRAE, Handbook-heating, Ventilating, and Air-conditioning Systems and
Equipment, American Society of Heating, Refrigerating and Air-Conditioning
Engineers, 2012.
[3] ASHRAE, Handbook-heating, Ventilating, and Air-conditioning Applications,
American Society of Heating, Refrigerating and Air-Conditioning Engineers,
2011.
[4] J. LeBrun, J.-P. Bourdouxhe, M. Grodent, HVAC1 Toolkit: A Toolkit for Primary
HVAC System Energy Calculation, American Society of Heating, Refrigerating
and Air-Conditioning Engineers, 1999.
[5] M.J. Brandemuehl, S. Gabel, I. Andresen, HVAC 2 Toolkit: A Toolkit for Sec-
ondary HVAC System Energy Calculations, American Society of Heating,
Refrigerating and Air-Conditioning Engineers, 1993.
[6] TRNSYS 16 A Transient Energy Simulation Program, Mathematical Reference,
Solar Energy Lab, University of Visconsin-Madison, 2005.
[7] EnergyPlus Engineering Reference: The Reference to EnergyPlus Calcula-
tions, DOE, 2013.
[8] G. Kelly, C. Park, D. Clark, W. May, HVACSIM, a dynamic building/HVAC/
control systems simulation program, in: Proc. Workshop HVAC Controls
Model. Simul., IBPSA, Seattle, Washington, United States, 1984, pp. 175e185.
[9] A.S. Kalagasidis, P. Weitzmann, T.R. Nielsen, R. Peuhkuri, C.-E. Hagentoft,
C. Rode, The international building physics toolbox in Simulink, Energy
Build. 39 (2007) 665e674.
[10] A. Husaunndee, R. Lahrech, H. Vaezi-Nejad, J. Visier, SIMBAD: a simulation
toolbox for the design and test of HVAC control systems, in: Proc. Build.
Simul., vol. 97IBPSA, Prague, Czech Republic, 1997, pp. 269e276.
[11] C. Wemhner, B. Hafner, K. Schwarzer, Simulation of solar thermal systems
with CARNOT blockset in the environment Matlab Simulink, in: Proc.
Eurosun. 2000 Conf., ISES, Copenhagen, Denmark, 2000.
[12] A.S. Kalagasidis, HAM-tools: an integrated simulation tool for heat, air and
moisture transfer analyses in building physics, in: Department of Building
Physics, Chalmers University of Technology Gothenburg, Sweden, 2004, p.
142.
[13] A. van Schijndel, J. Hensen, Integrated heat, air and moisture modeling
toolkit in Matlab, in: Proc. 9th Int. IBPSA Conf., IBPSA, Montreal, Canada,
2005, pp. 1107e1111.
[14] N. Mendes, G. Oliveira, H. Araujo, L. Coelho, A Matlab-based simulation tool
for building thermal performance analysis, in: Eighth Int. IBPSA Conf., IBPSA,
Eindhoven, Netherlands, 2003, pp. 855e862.
[15] P.R. Novak, N. Mendes, G.H. Oliveira, Simulation of HVAC plants in 2 Brazilian
cities using Matlab/Simulink, in: Ninth Int. IBPSA Conf., IBPSA, Montreal,
Canada, 2005, pp. 859e866.
[16] J. Singh, N. Singh, J.K. Sharma, Fuzzy modeling and control of HVAC systems
e a review, J. Sci. Ind. Res. 65 (2006) 470e476.
[17] M. Trcka, J.L.M. Hensen, Overview of HVAC system simulation, Autom.
Constr. 19 (2010) 93e99.
[18] J.A. Orosa, A.C. Oliveira, Software tools for HVAC research, Adv. Eng. Softw. 42
(2011) 846e851.
[19] P. Riederer, Matlab/Simulink for building and HVAC simulation-State of
the art, in: Ninth Int. IBPSA Conf., IBPSA, Montreal, Canada, 2005, pp. 1019e
1026.
[20] Q. Bi, W.-J. Cai, Q.-G. Wang, C.-C. Hang, E.-L. Lee, Y. Sun, K.-D. Liu, Y. Zhang,
B. Zou, Advanced controller auto-tuning and its application in HVAC systems,
Control Eng. Pract. 8 (2000) 633e644.
[21] J. Rehrl, M. Horn, Temperature control for HVAC systems based on exact
linearization and model predictive control, in: Int. Conf. Control Appl. (CCA),
IEEE, Denver, Colorado, United States, 2011, pp. 1119e1124.
[22] G. Huang, Model predictive control of VAV zone thermal systems concerning
bi-linearity and gain nonlinearity, Control Eng. Pract. 19 (2011) 700e710.
[23] G. Huang, S. Wang, X. Xu, Robust model predictive control of VAV air-
handling units concerning uncertainties and constraints, HVAC R. Res. 16
(2010) 15e33.
[24] S. Wang, X. Xu, G. Huang, Robust MPC for temperature control of air-
conditioning systems concerning on constraints and multitype un-
certainties, Build. Serv. Eng. Res. Technol. 31 (2010) 39e55.
[25] S.A. Kalogirou, Articial neural networks and genetic algorithms in energy
applications in buildings, Adv. Build. Energy Res. 3 (2009) 83e120.
[26] A. Kusiak, M. Li, Z. Zhang, A data-driven approach for steam load prediction
in buildings, Appl. Energy 87 (2010) 925e933.
[27] J. Chen, Z. Lian, L. Tan, W. Zhu, W. Zhang, Modeling and experimental
research on ground-source heat pump in operation by neural network, in:
Int. Conf. Comput. Distrib. Control Intell. Environ. Monit. (CDCIEM), IEEE,
Changsha, China, 2011, pp. 459e462.
[28] F. Tang, HVAC system modeling and optimization: a data-mining approach,
in: Department of Industrial Engineering, The University of Iowa, Iowa City,
Iowa, United States, 2010, p. 92.
[29] A. Kusiak, G. Xu, Modeling and optimization of HVAC systems using a dy-
namic neural network, Energy 42 (2012) 241e250.
[30] O. Morisot, D. Marchio, Fault detection and diagnosis on HVAC variable air
volume system using articial neural network, in: Proc. IBPSA Build. Simul.,
IBPSA, Kyoto, Japan, 1999, pp. 1027e1034.
[31] L. Ding, J. Lv, X. Li, L. Li, Support vector regression and ant colony optimi-
zation for HVAC cooling load prediction, in: Int. Symp. Comput. Commun.
Control. Autom. (3CA), vol. 1IEEE, Tainan, Taiwan, 2010, pp. 537e541.
[32] X. Li, L. Ding, Y. Deng, L. Li, Hybrid support vector machine and ARIMA model
in building cooling prediction, in: Int. Symp. Comput. Commun. Control.
Autom. (3CA), vol. 1IEEE, Tainan, Taiwan, 2010, pp. 533e536.
[33] X. Li, L. Ding, J. Lv, G. Xu, J. Li, A novel hybrid approach of KPCA and SVM for
building cooling load prediction, in: Third Int. Conf. Knowl. Discovery Data
Min. (WKDD), IEEE, Phuket, Thailand, 2010, pp. 522e526.
[34] X. Li, L. Ding, L. Li, A novel building cooling load prediction based on SVR and
SAPSO, in: Int. Symp. Comput. Commun. Control. Autom. (3CA), vol. 1IEEE,
Tainan, Taiwan, 2010, pp. 528e532.
[35] A. Pal, R. Mudi, Self-tuning fuzzy PI controller and its applications to HVAC
systems, Int. J. Comput. Cogn. 6 (2008) 25e30.
[36] K. Chen, Y. Jiao, E.S. Lee, Fuzzy adaptive networks in thermal comfort, Appl.
Math. Lett. 19 (2006) 420e426.
[37] H. Lu, L. Jia, S. Kong, Z. Zhang, Predictive functional control based on fuzzy Te
S model for HVAC systems temperature control, J. Control Theory Appl. 5
(2007) 94e98.
[38] R.Z. Homod, K.S.M. Sahari, H.A.F. Almurib, F.H. Nagi, RLF and TS fuzzy model
identication of indoor thermal comfort based on PMV/PPD, Build. Environ.
49 (2012) 141e153.
[39] S. Soyguder, H. Alli, Predicting of fan speed for energy saving in HVAC system
based on adaptive network based fuzzy inference system, Expert Syst. Appl.
36 (2009) 8631e8638.
[40] K. Li, H. Su, J. Chu, Forecasting building energy consumption using neural
networks and hybrid neuro-fuzzy system: a comparative study, Energy
Build. 43 (2011) 2893e2899.
[41] C.M.J. Yiu, Statistical modelling and forecasting schemes for air-conditioning
system, in: Department of Building Services Engineering, The Hong Kong
Polytechnic University, Hung Hom, Kowloon, Hong Kong, 2008, p. 181.
[42] A.A. Safa, Performance analysis of a two-stage variable capacity air source
heat pump and a horizontal loop coupled ground source heat pump system,
in: Mechanical and Industrial Engineering, Ryerson University, Toronto,
Canada, 2012, p. 130.
[43] B. Dong, Z. ONeill, D. Luo, B. Trevor, Development and calibration of a
reduced-order energy performance model for a mixed-use building, in: Proc.
BS2013: 13th Conf. Int. Build. Perform. Simul. Assoc., IBPSA, Chambery,
France, 2013, pp. 636e643.
[44] G.S. Virk, D.L. Loveday, Model-based control for HVAC applications, in: Proc.
Third IEEE Conf. Control. Appl., vol. 3IEEE, Glasgow, Scotland, 1994, pp.
1861e1866.
[45] J. Ma, S.J. Qin, B. Li, T. Salsbury, Economic model predictive control for
building energy systems, in: Innovative Smart Grid Technol. (ISGT), IEEE,
California, United States, 2011, pp. 1e6.
[46] J.-H. Huh, M.J. Brandemuehl, Optimization of air-conditioning system oper-
ating strategies for hot and humid climates, Energy Build. 40 (2008) 1202e
1213.
[47] Y.K. Penya, C.E. Borges, D. Agote, I. Fernandez, Short-term load forecasting in
air-conditioned non-residential Buildings, in: Int. Symp. Ind. Electron. (ISIE),
IEEE, Gdansk, Poland, 2011, pp. 1359e1364.
[48] L. Ferkl, J. Siroky, Ceiling radiant cooling: comparison of ARMAX and sub-
space identication modelling methods, Build. Environ. 45 (2010) 205e212.
[49] J. Nishiguchi, T. Konda, R. Dazai, Data-driven optimal control for building
energy conservation, in: Proc. SICE Annu. Conf., IEEE, Taipei, Taiwan, 2010,
pp. 116e120.
[50] I. Watson, A case study of maintenance of a commercially elded case-based
reasoning system, Comput Intell. 17 (2001) 387e398.
[51] I. Zlatanovic, K. Gligorevic, S. Ivanovic, N. Rudonja, Energy-saving estimation
model for hypermarket HVAC systems applications, Energy Build. 43 (2011)
3353e3359.
 A. Afram, F. Janabi-Shari / Applied Thermal Engineering 67 (2014) 507e519
[52] Y. Iino, M. Murai, D. Murayama, I. Motoyama, S. Kuzusaka, K. Ueta, Hybrid
modeling with physical and JIT model for building thermal load prediction
and optimal energy saving control, in: ICCAS-SICE, IEEE, Fukuoka, Japan,
2009, pp. 2008e2011.
[53] B. Tashtoush, M. Molhim, M. Al-Rousan, Dynamic model of an HVAC system
for control analysis, Energy 30 (2005) 1729e1745.
[54] S. Herzog, D. Atabay, J. Jungwirth, V. Mikulovic, Self-adapting building
models for model predictive control, in: Proc. BS2013: 13th Conf. Int. Build.
Perform. Simul. Assoc., IBPSA, Chambery, France, 2013, pp. 2489e2493.
[55] A. Thosar, A. Patra, S. Bhattacharyya, Feedback linearization based control of
a variable air volume air conditioning system for cooling applications, ISA
Trans. 47 (2008) 339e349.
[56] S. Wu, J.-Q. Sun, A physics-based linear parametric model of room temper-
ature in ofce buildings, Build. Environ. 50 (2012) 1e9.
[57] R.Z. Homod, Review on the HVAC system modeling types and the short-
comings of their application, J. Energy 2013 (2013) 1e10.
[58] Energy estimating and modeling methods, in: ASHRAE Handbook Funda-
mentals, American Society of Heating, Refrigerating and Air-Conditioning
Engineers, 2009.
[59] F. Tahersima, J. Stoustrup, H. Rasmussen, P.G. Nielsen, Thermal analysis of an
HVAC system with TRV controlled hydronic radiator, in: Conf. Autom. Sci.
Eng. (CASE), IEEE, Toronto, Ontario, Canada, 2010, pp. 756e761.
[60] C.O. Pedersen, A fundamental approach to cooling load calculations for
UFADS and stratied spaces, in: Eleventh Int. IBPSA Conf., IBPSA, Glasgow,
Scotland, 2009, pp. 704e708.
[61] M. Mossolly, K. Ghali, N. Ghaddar, Optimal control strategy for a multi-zone
air conditioning system using a genetic algorithm, Energy 34 (2009) 58e66.
[62] A. Beghi, L. Cecchinato, F. Paggiaro, M. Rampazzo, VAVAC systems modeling
and simulation for FDD applications, in: Int. Conf. Control. Autom. (ICCA),
IEEE, Santiago, Chile, 2011, pp. 800e805.
[63] I. Zajic, T. Larkowski, M. Sumislawska, K.J. Burnham, D. Hill, Modelling of an
air handling unit: a hammerstein-bilinear model identication approach, in:
21st Int. Conf. Syst. Eng. (ICSEng), IEEE, Las Vegas, Nevada, United States,
2011, pp. 59e63.
[64] A.W.M. van Schijndel, H.L. Schellen, W.J. Timmermans, Simulation of the
climate system performance of a museum in case of failure events, Energy
Build. 42 (2010) 1790e1796.
[65] F. Tahmasebi, R. Zach, M. Schub, A. Mahdavi, Simulation model calibration:
an optimization-based approach, in: Fourth German-Austrian IBPSA Conf.,
IBPSA, Berlin, Germany, 2012, pp. 386e391.
[66] J.A.F. Pascual, R.S. Garay, J.I. Torrens, M.M. Keane, Heat ventilation and air
conditioning modelling for model based fault detection and diagnosis, in:
Proc. BS2013: 13th Conf. Int. Build. Perform. Simul. Assoc., IBPSA, Chambery,
France, 2013, pp. 3513e3520.
[67] N. Matsushita, H. Yoshida, H. Suzuki, Application of the air-conditioning
system energy simulation for commissioning (ACSES/CX) tool to HVAC sys-
tem commissioning Part 2: application to the substation of a heat source
system with bleed-in control, in: 13th Conf. Int. Build. Performe Simul.
Assoc., IBPSA, Chambery, France, 2013, pp. 1778e1785.
[68] G.-Y. Jin, X.-D. Ding, P.-Y. Tan, T.-M. Koh, A hybrid water-cooled centrifugal
chiller model, in: Conf. Ind. Electron. Appl. (ICIEA), IEEE, Beijing, China, 2011,
pp. 2298e2303.
[69] A.P. Wemhoff, M.V. Frank, Predictions of energy savings in HVAC systems by
lumped models, Energy Build. 42 (2010) 1807e1814.
[70] P. Cui, H. Yang, J.D. Spitler, Z. Fang, Simulation of hybrid ground-coupled heat
pump with domestic hot water heating systems using HVACSIM, Energy
Build. 40 (2008) 1731e1736.
[71] X. Yan, Q. Ren, Q. Meng, Global optimization of VAV air conditioning system,
in: 8th World Congr. Intell. Control. Autom. (WCICA), IEEE, Jinan, China,
2010, pp. 5077e5081.
[72] B. Dong, Non-linear optimal controller design for building HVAC systems, in:
Int. Conf. Control. Appl. (CCA), IEEE, Yokohama, Japan, 2010, pp. 210e215.
[73] Y. Ma, F. Borrelli, B. Hencey, A. Packard, S. Bortoff, Model predictive control of
thermal energy storage in building cooling systems, in: Proc. 48th IEEE Conf.
Decis. Control., IEEE, Shanghai, China, 2009, pp. 392e397.
[74] S. Burhenne, J. Radon, M. Pazold, S. Herkel, F. Antretter, Integration of HVAC
models into a hygrothermal whole building simulation tool, in: Proc. of
Build. Simul. 2011: 12th Conf. Int. Build. Perform. Simul. Assoc., IBPSA,
Sydney, Australia, 2011, pp. 1777e1783.
[75] Energy estimating and modeling methods, in: Handbook Fundamentals,
American Society of Heating, Refrigerating and Air-Conditioning Engineers,
2009, pp. 19.11e19.39.
519
[76] P. Mi-Hyun, S. Eun-Gyeong, L. Heung-Reol, S. In-Soo, Dynamic model and
control algorithm of HVAC system for OLEV application, in: Int. Conf.
Control. Autom. Syst. (ICCAS), IEEE, Gyeonggi-do, Korea, 2010, pp. 1312e
1317.
[77] P. Sreedharan, P. Haves, Comparison of chiller models for use in model-based
fault detection, in: Int. Conf. Enhanced Build. Oper., Austin, Texas, United
States, 2001, pp. 1e10.
[78] Y. Yuan, X. Cao, L. Sun, B. Lei, N. Yu, Ground source heat pump system: a
review of simulation in China, Renew. Sustain. Energy Rev. 16 (2012) 6814.
[79] X. Xu, Simulation and Optimal Control of Hybrid Ground Source Heat Pump
Systems, Oklahoma State University, Stillwater, Oklahoma, USA, 2007, p. 395.
[80] H. Jin, Parameter Estimation Based Models of Water Source Heat Pumps,
Oklahoma State University, Stillwater, Oklahoma, USA, 2002, p. 382.
[81] G.-Y. Jin, W.-J. Cai, L. Lu, E.L. Lee, A. Chiang, A simplied modeling of me-
chanical cooling tower for control and optimization of HVAC systems, Energy
Convers. Manag. 48 (2007) 355e365.
[82] V. Vakiloroaya, B. Samali, J. Madadnia, Q.P. Ha, Component-wise optimization
for a commercial central cooling plant, in: 37th Ann. Conf. IEEE Ind. Electron.
Soc. (IECON), IEEE, Melbourne, Victoria, Australia, 2011, pp. 2769e2774.
[83] J. Li, G. Poulton, G. Platt, J. Wall, G. James, Dynamic zone modelling for HVAC
system control, Int. J. Model. Identif. Control 9 (2010) 5e13.
[84] R. Balan, J. Cooper, K.-M. Chao, S. Stan, R. Donca, Parameter identication and
model based predictive control of temperature inside a house, Energy Build.
43 (2011) 748e758.
[85] J. Leclere, F. Wurtz, E. Wurtz, A low order envelope model for optimised
predictive control of indoor temperature: development methodology and
calibration with a numerical model, in: Proc. BS2013: 13th Conf. Int. Build.
Perform. Simul. Assoc., IBPSA, Chambery, France, 2013, pp. 2341e2348.
[86] J. Liyan, Y. Zhe, M.C. Murphy-Hoye, A. Pratt, E.G. Piccioli, T. Lang, Multi-scale
stochastic optimization for home energy management, in: 4th IEEE Int.
Workshop Comput. Adv. Multi-sens. Adapt. Process (CAMSAP), IEEE, San
Juan, Puerto Rico, 2011, pp. 113e116.
[87] N. Hariharan, B.P. Rasmussen, Parameter estimation for dynamic HVAC
models with limited sensor information, in: Am. Control. Conf. (ACC), IEEE,
Baltimore, Maryland, United States, 2010, pp. 5886e5891.
[88] G.-Y. Jin, P.-Y. Tan, X.-D. Ding, T.-M. Koh, Cooling coil unit dynamic control of
in HVAC system, in: Conf. Ind. Electron. Appl. (ICIEA), IEEE, Beijing, China,
2011, pp. 942e947.
[89] G.-Y. Jin, W.-J. Cai, Y.-W. Wang, Y. Yao, A simple dynamic model of cooling
coil unit, Energy Convers. Manag. 47 (2006) 2659e2672.
[90] J. Berardino, C. Nwankpa, Dynamic load modeling of an HVAC chiller for
demand response applications, in: First IEEE Int. Conf. Smart. Grid Commun.
(SmartGridComm), IEEE, Gaithersburg, Maryland, United States, 2010, pp.
108e113.
[91] M. Kumar, I.N. Kar, Non-linear HVAC computations using least square sup-
port vector machines, Energy Convers. Manag. 50 (2009) 1411e1418.
[92] G. Mustafaraj, J. Chen, G. Lowry, Thermal behaviour prediction utilizing
articial neural networks for an open ofce, Appl. Math. Model. 34 (2010)
3216e3230.
[93] G.P. Henze, C. Felsmann, D.E. Kalz, G. Knabe, Impact of forecasting accuracy
on predictive optimal control of active and passive building thermal storage
inventory, HVAC R. Res. 10 (2004) 153e178.
[94] F. Tahmasebi, A. Mahdavi, A two-staged simulation model calibration
approach to virtual sensors for building performance data, in: Proc. BS2013:
13th Conf. Int. Build. Perform. Simul. Assoc., IBPSA, Chambery, France, 2013,
pp. 608e613.
[95] G. Casella, R.L. Berger, Statistical Inference, Duxbury Press Belmont, CA, 1990.
[96] ASHRAE, Standard 140-2011 Standard Method of Test for the Evaluation of
Building Energy Analysis Computer Programs, 2011.
[97] J. Ron, N. Joel, Model validation and testing: the methodological foundation
of ASHRAE standard 140, in: ASHRAE 2006 Annu. Meeting, vol. 112ASHRAE,
Quebec City, Canada, 2006, pp. 367e379.
[98] R. Judkoff, J. Neymark, Home energy Rating System Building Energy Simu-
lation Test (HERS BESTEST), in: Tier 1 and Tier 2 Tests Users Manual, vol. 1,
NREL, Golden, Colorado, USA, 1995.
[99] ISO, ISO 13792:2012 Thermal Performance of Buildings e Calculation of In-
ternal Temperatures of a Room in Summer without Mechanical Cooling e
Simplied Methods, 2012.
[100] ISO, ISO 13791:2012 Thermal Performance of Buildings e Calculation of
Internal Temperatures of a Room in Summer without Mechanical Cooling e
General Criteria and Validation Procedures, 2012.